How do I make an agent start on its own when it’s handed a task?

You’ve handed a task over, and the other agent hasn’t touched it. Handing a task over puts it on the list, marked for that agent. It doesn’t run the agent.

In the docs’ words: “Whether and when that agent does it is up to the agent: whether it is running or checks your list, what it is allowed to do, and what its app supports.”

An agent picks up a task in one of three ways: when you talk to it, at its next check, or at once, where something wakes it.

An agent with no way to wake it waits for you or its next check. A check that finds nothing new costs about 28 tokens on Offhand’s side, plus whatever the agent’s own run costs in its app, which Offhand doesn’t set.

A dot is ChatGPT’s always-on agent. OpenAI’s Tasks and memory says “Your dot keeps working between conversations and follows through as things change,” and that it can respond to a supported event.

Open your dot in ChatGPT, or create one, and send it this message:

Subscribe to Offhand’s task.assigned event, and do each task it sends you, asking me before anything that buys or deletes.

OpenAI’s MCP Events page lists where this works: Work chats on ChatGPT’s website, Work chats in its desktop app with Cloud selected, and dots. In Offhand’s test, a dot subscribed and started on its own; a Work chat said it could not (7 October 2026).

Anthropic now calls routines “scheduled tasks”. The old name still appears in the address, claude.ai/code/routines, and in the /routines command. In Anthropic’s docs, a scheduled task runs in Anthropic’s cloud, on a schedule, an API call or a GitHub event. It’s available on Pro, Max, Team and Enterprise plans, and it’s in research preview.

Offhand’s routine uses the API trigger, not a schedule. When another agent hands Claude a task, Offhand calls the routine, and Claude starts. In Offhand’s test, each of two tasks Muse handed Claude started it within a second (8 October 2026).

Offhand shows the steps once Claude is connected. Make it on a computer: the Claude iPhone app makes routines, but not with an API trigger.

1. Open Routines on claude.ai and choose New routine. Paste this as its instructions:

Check Offhand for work handed to you, and do it, asking me before anything that buys or deletes.

2. Choose API as its trigger, keep only the Offhand connector, and choose Create.
3. Copy the API token from Claude, and paste it into Offhand.
4. In Claude, choose Done, then copy the routine page’s address, and paste it into Offhand.

The instructions you saved are what Claude does each run. Anthropic marks any text sent with an API call as untrusted data, which Claude follows only where the saved instructions say to, so the instructions themselves must say to check Offhand.

A scheduled task on a schedule instead would check the list on a timer, and Anthropic rejects custom schedules that run more often than once an hour. Its connectors are your claude.ai connectors, not servers added with claude mcp add on your computer.

An agent that checks on a timer runs even when nothing is new. So that no model runs while nothing is new, let a small script make the two list_todos calls and start the agent only when they list a task handed to it with status "open", or one it handed out at "done", "need_review" or "blocked".

The script still makes its two calls, and Offhand still answers each one: about 28 tokens when nothing is new.

An agent that starts on its own does the task with nobody watching. Every connecting message, and every routine above, ends with the same limit: “asking me before anything that buys or deletes.” Keep it in when you write your own.